1;
2; jiss2fst.asm - fast integer IDCT (SSE2)
3;
4; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
5;
6; Based on
7; x86 SIMD extension for IJG JPEG library
8; Copyright (C) 1999-2006, MIYASAKA Masaru.
9; For conditions of distribution and use, see copyright notice in jsimdext.inc
10;
11; This file should be assembled with NASM (Netwide Assembler),
12; can *not* be assembled with Microsoft's MASM or any compatible
13; assembler (including Borland's Turbo Assembler).
14; NASM is available from http://nasm.sourceforge.net/ or
15; http://sourceforge.net/project/showfiles.php?group_id=6208
16;
17; This file contains a fast, not so accurate integer implementation of
18; the inverse DCT (Discrete Cosine Transform). The following code is
19; based directly on the IJG's original jidctfst.c; see the jidctfst.c
20; for more details.
21;
22; [TAB8]
23
24%include "jsimdext.inc"
25%include "jdct.inc"
26
27; --------------------------------------------------------------------------
28
29%define CONST_BITS	8	; 14 is also OK.
30%define PASS1_BITS	2
31
32%if IFAST_SCALE_BITS != PASS1_BITS
33%error "'IFAST_SCALE_BITS' must be equal to 'PASS1_BITS'."
34%endif
35
36%if CONST_BITS == 8
37F_1_082	equ	277		; FIX(1.082392200)
38F_1_414	equ	362		; FIX(1.414213562)
39F_1_847	equ	473		; FIX(1.847759065)
40F_2_613	equ	669		; FIX(2.613125930)
41F_1_613	equ	(F_2_613 - 256)	; FIX(2.613125930) - FIX(1)
42%else
43; NASM cannot do compile-time arithmetic on floating-point constants.
44%define	DESCALE(x,n)  (((x)+(1<<((n)-1)))>>(n))
45F_1_082	equ	DESCALE(1162209775,30-CONST_BITS)	; FIX(1.082392200)
46F_1_414	equ	DESCALE(1518500249,30-CONST_BITS)	; FIX(1.414213562)
47F_1_847	equ	DESCALE(1984016188,30-CONST_BITS)	; FIX(1.847759065)
48F_2_613	equ	DESCALE(2805822602,30-CONST_BITS)	; FIX(2.613125930)
49F_1_613	equ	(F_2_613 - (1 << CONST_BITS))	; FIX(2.613125930) - FIX(1)
50%endif
51
52; --------------------------------------------------------------------------
53	SECTION	SEG_CONST
54
55; PRE_MULTIPLY_SCALE_BITS <= 2 (to avoid overflow)
56; CONST_BITS + CONST_SHIFT + PRE_MULTIPLY_SCALE_BITS == 16 (for pmulhw)
57
58%define PRE_MULTIPLY_SCALE_BITS   2
59%define CONST_SHIFT     (16 - PRE_MULTIPLY_SCALE_BITS - CONST_BITS)
60
61	alignz	16
62	global	EXTN(jconst_idct_ifast_sse2) PRIVATE
63
64EXTN(jconst_idct_ifast_sse2):
65
66PW_F1414	times 8 dw  F_1_414 << CONST_SHIFT
67PW_F1847	times 8 dw  F_1_847 << CONST_SHIFT
68PW_MF1613	times 8 dw -F_1_613 << CONST_SHIFT
69PW_F1082	times 8 dw  F_1_082 << CONST_SHIFT
70PB_CENTERJSAMP	times 16 db CENTERJSAMPLE
71
72	alignz	16
73
74; --------------------------------------------------------------------------
75	SECTION	SEG_TEXT
76	BITS	32
77;
78; Perform dequantization and inverse DCT on one block of coefficients.
79;
80; GLOBAL(void)
81; jsimd_idct_ifast_sse2 (void * dct_table, JCOEFPTR coef_block,
82;                       JSAMPARRAY output_buf, JDIMENSION output_col)
83;
84
85%define dct_table(b)	(b)+8			; jpeg_component_info * compptr
86%define coef_block(b)	(b)+12		; JCOEFPTR coef_block
87%define output_buf(b)	(b)+16		; JSAMPARRAY output_buf
88%define output_col(b)	(b)+20		; JDIMENSION output_col
89
90%define original_ebp	ebp+0
91%define wk(i)		ebp-(WK_NUM-(i))*SIZEOF_XMMWORD	; xmmword wk[WK_NUM]
92%define WK_NUM		2
93
94	align	16
95	global	EXTN(jsimd_idct_ifast_sse2) PRIVATE
96
97EXTN(jsimd_idct_ifast_sse2):
98	push	ebp
99	mov	eax,esp				; eax = original ebp
100	sub	esp, byte 4
101	and	esp, byte (-SIZEOF_XMMWORD)	; align to 128 bits
102	mov	[esp],eax
103	mov	ebp,esp				; ebp = aligned ebp
104	lea	esp, [wk(0)]
105	pushpic	ebx
106;	push	ecx		; unused
107;	push	edx		; need not be preserved
108	push	esi
109	push	edi
110
111	get_GOT	ebx		; get GOT address
112
113	; ---- Pass 1: process columns from input.
114
115;	mov	eax, [original_ebp]
116	mov	edx, POINTER [dct_table(eax)]	; quantptr
117	mov	esi, JCOEFPTR [coef_block(eax)]		; inptr
118
119%ifndef NO_ZERO_COLUMN_TEST_IFAST_SSE2
120	mov	eax, DWORD [DWBLOCK(1,0,esi,SIZEOF_JCOEF)]
121	or	eax, DWORD [DWBLOCK(2,0,esi,SIZEOF_JCOEF)]
122	jnz	near .columnDCT
123
124	movdqa	xmm0, XMMWORD [XMMBLOCK(1,0,esi,SIZEOF_JCOEF)]
125	movdqa	xmm1, XMMWORD [XMMBLOCK(2,0,esi,SIZEOF_JCOEF)]
126	por	xmm0, XMMWORD [XMMBLOCK(3,0,esi,SIZEOF_JCOEF)]
127	por	xmm1, XMMWORD [XMMBLOCK(4,0,esi,SIZEOF_JCOEF)]
128	por	xmm0, XMMWORD [XMMBLOCK(5,0,esi,SIZEOF_JCOEF)]
129	por	xmm1, XMMWORD [XMMBLOCK(6,0,esi,SIZEOF_JCOEF)]
130	por	xmm0, XMMWORD [XMMBLOCK(7,0,esi,SIZEOF_JCOEF)]
131	por	xmm1,xmm0
132	packsswb xmm1,xmm1
133	packsswb xmm1,xmm1
134	movd	eax,xmm1
135	test	eax,eax
136	jnz	short .columnDCT
137
138	; -- AC terms all zero
139
140	movdqa	xmm0, XMMWORD [XMMBLOCK(0,0,esi,SIZEOF_JCOEF)]
141	pmullw	xmm0, XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
142
143	movdqa    xmm7,xmm0		; xmm0=in0=(00 01 02 03 04 05 06 07)
144	punpcklwd xmm0,xmm0		; xmm0=(00 00 01 01 02 02 03 03)
145	punpckhwd xmm7,xmm7		; xmm7=(04 04 05 05 06 06 07 07)
146
147	pshufd	xmm6,xmm0,0x00		; xmm6=col0=(00 00 00 00 00 00 00 00)
148	pshufd	xmm2,xmm0,0x55		; xmm2=col1=(01 01 01 01 01 01 01 01)
149	pshufd	xmm5,xmm0,0xAA		; xmm5=col2=(02 02 02 02 02 02 02 02)
150	pshufd	xmm0,xmm0,0xFF		; xmm0=col3=(03 03 03 03 03 03 03 03)
151	pshufd	xmm1,xmm7,0x00		; xmm1=col4=(04 04 04 04 04 04 04 04)
152	pshufd	xmm4,xmm7,0x55		; xmm4=col5=(05 05 05 05 05 05 05 05)
153	pshufd	xmm3,xmm7,0xAA		; xmm3=col6=(06 06 06 06 06 06 06 06)
154	pshufd	xmm7,xmm7,0xFF		; xmm7=col7=(07 07 07 07 07 07 07 07)
155
156	movdqa	XMMWORD [wk(0)], xmm2	; wk(0)=col1
157	movdqa	XMMWORD [wk(1)], xmm0	; wk(1)=col3
158	jmp	near .column_end
159	alignx	16,7
160%endif
161.columnDCT:
162
163	; -- Even part
164
165	movdqa	xmm0, XMMWORD [XMMBLOCK(0,0,esi,SIZEOF_JCOEF)]
166	movdqa	xmm1, XMMWORD [XMMBLOCK(2,0,esi,SIZEOF_JCOEF)]
167	pmullw	xmm0, XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_IFAST_MULT_TYPE)]
168	pmullw	xmm1, XMMWORD [XMMBLOCK(2,0,edx,SIZEOF_IFAST_MULT_TYPE)]
169	movdqa	xmm2, XMMWORD [XMMBLOCK(4,0,esi,SIZEOF_JCOEF)]
170	movdqa	xmm3, XMMWORD [XMMBLOCK(6,0,esi,SIZEOF_JCOEF)]
171	pmullw	xmm2, XMMWORD [XMMBLOCK(4,0,edx,SIZEOF_IFAST_MULT_TYPE)]
172	pmullw	xmm3, XMMWORD [XMMBLOCK(6,0,edx,SIZEOF_IFAST_MULT_TYPE)]
173
174	movdqa	xmm4,xmm0
175	movdqa	xmm5,xmm1
176	psubw	xmm0,xmm2		; xmm0=tmp11
177	psubw	xmm1,xmm3
178	paddw	xmm4,xmm2		; xmm4=tmp10
179	paddw	xmm5,xmm3		; xmm5=tmp13
180
181	psllw	xmm1,PRE_MULTIPLY_SCALE_BITS
182	pmulhw	xmm1,[GOTOFF(ebx,PW_F1414)]
183	psubw	xmm1,xmm5		; xmm1=tmp12
184
185	movdqa	xmm6,xmm4
186	movdqa	xmm7,xmm0
187	psubw	xmm4,xmm5		; xmm4=tmp3
188	psubw	xmm0,xmm1		; xmm0=tmp2
189	paddw	xmm6,xmm5		; xmm6=tmp0
190	paddw	xmm7,xmm1		; xmm7=tmp1
191
192	movdqa	XMMWORD [wk(1)], xmm4	; wk(1)=tmp3
193	movdqa	XMMWORD [wk(0)], xmm0	; wk(0)=tmp2
194
195	; -- Odd part
196
197	movdqa	xmm2, XMMWORD [XMMBLOCK(1,0,esi,SIZEOF_JCOEF)]
198	movdqa	xmm3, XMMWORD [XMMBLOCK(3,0,esi,SIZEOF_JCOEF)]
199	pmullw	xmm2, XMMWORD [XMMBLOCK(1,0,edx,SIZEOF_IFAST_MULT_TYPE)]
200	pmullw	xmm3, XMMWORD [XMMBLOCK(3,0,edx,SIZEOF_IFAST_MULT_TYPE)]
201	movdqa	xmm5, XMMWORD [XMMBLOCK(5,0,esi,SIZEOF_JCOEF)]
202	movdqa	xmm1, XMMWORD [XMMBLOCK(7,0,esi,SIZEOF_JCOEF)]
203	pmullw	xmm5, XMMWORD [XMMBLOCK(5,0,edx,SIZEOF_IFAST_MULT_TYPE)]
204	pmullw	xmm1, XMMWORD [XMMBLOCK(7,0,edx,SIZEOF_IFAST_MULT_TYPE)]
205
206	movdqa	xmm4,xmm2
207	movdqa	xmm0,xmm5
208	psubw	xmm2,xmm1		; xmm2=z12
209	psubw	xmm5,xmm3		; xmm5=z10
210	paddw	xmm4,xmm1		; xmm4=z11
211	paddw	xmm0,xmm3		; xmm0=z13
212
213	movdqa	xmm1,xmm5		; xmm1=z10(unscaled)
214	psllw	xmm2,PRE_MULTIPLY_SCALE_BITS
215	psllw	xmm5,PRE_MULTIPLY_SCALE_BITS
216
217	movdqa	xmm3,xmm4
218	psubw	xmm4,xmm0
219	paddw	xmm3,xmm0		; xmm3=tmp7
220
221	psllw	xmm4,PRE_MULTIPLY_SCALE_BITS
222	pmulhw	xmm4,[GOTOFF(ebx,PW_F1414)]	; xmm4=tmp11
223
224	; To avoid overflow...
225	;
226	; (Original)
227	; tmp12 = -2.613125930 * z10 + z5;
228	;
229	; (This implementation)
230	; tmp12 = (-1.613125930 - 1) * z10 + z5;
231	;       = -1.613125930 * z10 - z10 + z5;
232
233	movdqa	xmm0,xmm5
234	paddw	xmm5,xmm2
235	pmulhw	xmm5,[GOTOFF(ebx,PW_F1847)]	; xmm5=z5
236	pmulhw	xmm0,[GOTOFF(ebx,PW_MF1613)]
237	pmulhw	xmm2,[GOTOFF(ebx,PW_F1082)]
238	psubw	xmm0,xmm1
239	psubw	xmm2,xmm5		; xmm2=tmp10
240	paddw	xmm0,xmm5		; xmm0=tmp12
241
242	; -- Final output stage
243
244	psubw	xmm0,xmm3		; xmm0=tmp6
245	movdqa	xmm1,xmm6
246	movdqa	xmm5,xmm7
247	paddw	xmm6,xmm3		; xmm6=data0=(00 01 02 03 04 05 06 07)
248	paddw	xmm7,xmm0		; xmm7=data1=(10 11 12 13 14 15 16 17)
249	psubw	xmm1,xmm3		; xmm1=data7=(70 71 72 73 74 75 76 77)
250	psubw	xmm5,xmm0		; xmm5=data6=(60 61 62 63 64 65 66 67)
251	psubw	xmm4,xmm0		; xmm4=tmp5
252
253	movdqa    xmm3,xmm6		; transpose coefficients(phase 1)
254	punpcklwd xmm6,xmm7		; xmm6=(00 10 01 11 02 12 03 13)
255	punpckhwd xmm3,xmm7		; xmm3=(04 14 05 15 06 16 07 17)
256	movdqa    xmm0,xmm5		; transpose coefficients(phase 1)
257	punpcklwd xmm5,xmm1		; xmm5=(60 70 61 71 62 72 63 73)
258	punpckhwd xmm0,xmm1		; xmm0=(64 74 65 75 66 76 67 77)
259
260	movdqa	xmm7, XMMWORD [wk(0)]	; xmm7=tmp2
261	movdqa	xmm1, XMMWORD [wk(1)]	; xmm1=tmp3
262
263	movdqa	XMMWORD [wk(0)], xmm5	; wk(0)=(60 70 61 71 62 72 63 73)
264	movdqa	XMMWORD [wk(1)], xmm0	; wk(1)=(64 74 65 75 66 76 67 77)
265
266	paddw	xmm2,xmm4		; xmm2=tmp4
267	movdqa	xmm5,xmm7
268	movdqa	xmm0,xmm1
269	paddw	xmm7,xmm4		; xmm7=data2=(20 21 22 23 24 25 26 27)
270	paddw	xmm1,xmm2		; xmm1=data4=(40 41 42 43 44 45 46 47)
271	psubw	xmm5,xmm4		; xmm5=data5=(50 51 52 53 54 55 56 57)
272	psubw	xmm0,xmm2		; xmm0=data3=(30 31 32 33 34 35 36 37)
273
274	movdqa    xmm4,xmm7		; transpose coefficients(phase 1)
275	punpcklwd xmm7,xmm0		; xmm7=(20 30 21 31 22 32 23 33)
276	punpckhwd xmm4,xmm0		; xmm4=(24 34 25 35 26 36 27 37)
277	movdqa    xmm2,xmm1		; transpose coefficients(phase 1)
278	punpcklwd xmm1,xmm5		; xmm1=(40 50 41 51 42 52 43 53)
279	punpckhwd xmm2,xmm5		; xmm2=(44 54 45 55 46 56 47 57)
280
281	movdqa    xmm0,xmm3		; transpose coefficients(phase 2)
282	punpckldq xmm3,xmm4		; xmm3=(04 14 24 34 05 15 25 35)
283	punpckhdq xmm0,xmm4		; xmm0=(06 16 26 36 07 17 27 37)
284	movdqa    xmm5,xmm6		; transpose coefficients(phase 2)
285	punpckldq xmm6,xmm7		; xmm6=(00 10 20 30 01 11 21 31)
286	punpckhdq xmm5,xmm7		; xmm5=(02 12 22 32 03 13 23 33)
287
288	movdqa	xmm4, XMMWORD [wk(0)]	; xmm4=(60 70 61 71 62 72 63 73)
289	movdqa	xmm7, XMMWORD [wk(1)]	; xmm7=(64 74 65 75 66 76 67 77)
290
291	movdqa	XMMWORD [wk(0)], xmm3	; wk(0)=(04 14 24 34 05 15 25 35)
292	movdqa	XMMWORD [wk(1)], xmm0	; wk(1)=(06 16 26 36 07 17 27 37)
293
294	movdqa    xmm3,xmm1		; transpose coefficients(phase 2)
295	punpckldq xmm1,xmm4		; xmm1=(40 50 60 70 41 51 61 71)
296	punpckhdq xmm3,xmm4		; xmm3=(42 52 62 72 43 53 63 73)
297	movdqa    xmm0,xmm2		; transpose coefficients(phase 2)
298	punpckldq xmm2,xmm7		; xmm2=(44 54 64 74 45 55 65 75)
299	punpckhdq xmm0,xmm7		; xmm0=(46 56 66 76 47 57 67 77)
300
301	movdqa     xmm4,xmm6		; transpose coefficients(phase 3)
302	punpcklqdq xmm6,xmm1		; xmm6=col0=(00 10 20 30 40 50 60 70)
303	punpckhqdq xmm4,xmm1		; xmm4=col1=(01 11 21 31 41 51 61 71)
304	movdqa     xmm7,xmm5		; transpose coefficients(phase 3)
305	punpcklqdq xmm5,xmm3		; xmm5=col2=(02 12 22 32 42 52 62 72)
306	punpckhqdq xmm7,xmm3		; xmm7=col3=(03 13 23 33 43 53 63 73)
307
308	movdqa	xmm1, XMMWORD [wk(0)]	; xmm1=(04 14 24 34 05 15 25 35)
309	movdqa	xmm3, XMMWORD [wk(1)]	; xmm3=(06 16 26 36 07 17 27 37)
310
311	movdqa	XMMWORD [wk(0)], xmm4	; wk(0)=col1
312	movdqa	XMMWORD [wk(1)], xmm7	; wk(1)=col3
313
314	movdqa     xmm4,xmm1		; transpose coefficients(phase 3)
315	punpcklqdq xmm1,xmm2		; xmm1=col4=(04 14 24 34 44 54 64 74)
316	punpckhqdq xmm4,xmm2		; xmm4=col5=(05 15 25 35 45 55 65 75)
317	movdqa     xmm7,xmm3		; transpose coefficients(phase 3)
318	punpcklqdq xmm3,xmm0		; xmm3=col6=(06 16 26 36 46 56 66 76)
319	punpckhqdq xmm7,xmm0		; xmm7=col7=(07 17 27 37 47 57 67 77)
320.column_end:
321
322	; -- Prefetch the next coefficient block
323
324	prefetchnta [esi + DCTSIZE2*SIZEOF_JCOEF + 0*32]
325	prefetchnta [esi + DCTSIZE2*SIZEOF_JCOEF + 1*32]
326	prefetchnta [esi + DCTSIZE2*SIZEOF_JCOEF + 2*32]
327	prefetchnta [esi + DCTSIZE2*SIZEOF_JCOEF + 3*32]
328
329	; ---- Pass 2: process rows from work array, store into output array.
330
331	mov	eax, [original_ebp]
332	mov	edi, JSAMPARRAY [output_buf(eax)]	; (JSAMPROW *)
333	mov	eax, JDIMENSION [output_col(eax)]
334
335	; -- Even part
336
337	; xmm6=col0, xmm5=col2, xmm1=col4, xmm3=col6
338
339	movdqa	xmm2,xmm6
340	movdqa	xmm0,xmm5
341	psubw	xmm6,xmm1		; xmm6=tmp11
342	psubw	xmm5,xmm3
343	paddw	xmm2,xmm1		; xmm2=tmp10
344	paddw	xmm0,xmm3		; xmm0=tmp13
345
346	psllw	xmm5,PRE_MULTIPLY_SCALE_BITS
347	pmulhw	xmm5,[GOTOFF(ebx,PW_F1414)]
348	psubw	xmm5,xmm0		; xmm5=tmp12
349
350	movdqa	xmm1,xmm2
351	movdqa	xmm3,xmm6
352	psubw	xmm2,xmm0		; xmm2=tmp3
353	psubw	xmm6,xmm5		; xmm6=tmp2
354	paddw	xmm1,xmm0		; xmm1=tmp0
355	paddw	xmm3,xmm5		; xmm3=tmp1
356
357	movdqa	xmm0, XMMWORD [wk(0)]	; xmm0=col1
358	movdqa	xmm5, XMMWORD [wk(1)]	; xmm5=col3
359
360	movdqa	XMMWORD [wk(0)], xmm2	; wk(0)=tmp3
361	movdqa	XMMWORD [wk(1)], xmm6	; wk(1)=tmp2
362
363	; -- Odd part
364
365	; xmm0=col1, xmm5=col3, xmm4=col5, xmm7=col7
366
367	movdqa	xmm2,xmm0
368	movdqa	xmm6,xmm4
369	psubw	xmm0,xmm7		; xmm0=z12
370	psubw	xmm4,xmm5		; xmm4=z10
371	paddw	xmm2,xmm7		; xmm2=z11
372	paddw	xmm6,xmm5		; xmm6=z13
373
374	movdqa	xmm7,xmm4		; xmm7=z10(unscaled)
375	psllw	xmm0,PRE_MULTIPLY_SCALE_BITS
376	psllw	xmm4,PRE_MULTIPLY_SCALE_BITS
377
378	movdqa	xmm5,xmm2
379	psubw	xmm2,xmm6
380	paddw	xmm5,xmm6		; xmm5=tmp7
381
382	psllw	xmm2,PRE_MULTIPLY_SCALE_BITS
383	pmulhw	xmm2,[GOTOFF(ebx,PW_F1414)]	; xmm2=tmp11
384
385	; To avoid overflow...
386	;
387	; (Original)
388	; tmp12 = -2.613125930 * z10 + z5;
389	;
390	; (This implementation)
391	; tmp12 = (-1.613125930 - 1) * z10 + z5;
392	;       = -1.613125930 * z10 - z10 + z5;
393
394	movdqa	xmm6,xmm4
395	paddw	xmm4,xmm0
396	pmulhw	xmm4,[GOTOFF(ebx,PW_F1847)]	; xmm4=z5
397	pmulhw	xmm6,[GOTOFF(ebx,PW_MF1613)]
398	pmulhw	xmm0,[GOTOFF(ebx,PW_F1082)]
399	psubw	xmm6,xmm7
400	psubw	xmm0,xmm4		; xmm0=tmp10
401	paddw	xmm6,xmm4		; xmm6=tmp12
402
403	; -- Final output stage
404
405	psubw	xmm6,xmm5		; xmm6=tmp6
406	movdqa	xmm7,xmm1
407	movdqa	xmm4,xmm3
408	paddw	xmm1,xmm5		; xmm1=data0=(00 10 20 30 40 50 60 70)
409	paddw	xmm3,xmm6		; xmm3=data1=(01 11 21 31 41 51 61 71)
410	psraw	xmm1,(PASS1_BITS+3)	; descale
411	psraw	xmm3,(PASS1_BITS+3)	; descale
412	psubw	xmm7,xmm5		; xmm7=data7=(07 17 27 37 47 57 67 77)
413	psubw	xmm4,xmm6		; xmm4=data6=(06 16 26 36 46 56 66 76)
414	psraw	xmm7,(PASS1_BITS+3)	; descale
415	psraw	xmm4,(PASS1_BITS+3)	; descale
416	psubw	xmm2,xmm6		; xmm2=tmp5
417
418	packsswb  xmm1,xmm4	; xmm1=(00 10 20 30 40 50 60 70 06 16 26 36 46 56 66 76)
419	packsswb  xmm3,xmm7	; xmm3=(01 11 21 31 41 51 61 71 07 17 27 37 47 57 67 77)
420
421	movdqa	xmm5, XMMWORD [wk(1)]	; xmm5=tmp2
422	movdqa	xmm6, XMMWORD [wk(0)]	; xmm6=tmp3
423
424	paddw	xmm0,xmm2		; xmm0=tmp4
425	movdqa	xmm4,xmm5
426	movdqa	xmm7,xmm6
427	paddw	xmm5,xmm2		; xmm5=data2=(02 12 22 32 42 52 62 72)
428	paddw	xmm6,xmm0		; xmm6=data4=(04 14 24 34 44 54 64 74)
429	psraw	xmm5,(PASS1_BITS+3)	; descale
430	psraw	xmm6,(PASS1_BITS+3)	; descale
431	psubw	xmm4,xmm2		; xmm4=data5=(05 15 25 35 45 55 65 75)
432	psubw	xmm7,xmm0		; xmm7=data3=(03 13 23 33 43 53 63 73)
433	psraw	xmm4,(PASS1_BITS+3)	; descale
434	psraw	xmm7,(PASS1_BITS+3)	; descale
435
436	movdqa    xmm2,[GOTOFF(ebx,PB_CENTERJSAMP)]	; xmm2=[PB_CENTERJSAMP]
437
438	packsswb  xmm5,xmm6	; xmm5=(02 12 22 32 42 52 62 72 04 14 24 34 44 54 64 74)
439	packsswb  xmm7,xmm4	; xmm7=(03 13 23 33 43 53 63 73 05 15 25 35 45 55 65 75)
440
441	paddb     xmm1,xmm2
442	paddb     xmm3,xmm2
443	paddb     xmm5,xmm2
444	paddb     xmm7,xmm2
445
446	movdqa    xmm0,xmm1	; transpose coefficients(phase 1)
447	punpcklbw xmm1,xmm3	; xmm1=(00 01 10 11 20 21 30 31 40 41 50 51 60 61 70 71)
448	punpckhbw xmm0,xmm3	; xmm0=(06 07 16 17 26 27 36 37 46 47 56 57 66 67 76 77)
449	movdqa    xmm6,xmm5	; transpose coefficients(phase 1)
450	punpcklbw xmm5,xmm7	; xmm5=(02 03 12 13 22 23 32 33 42 43 52 53 62 63 72 73)
451	punpckhbw xmm6,xmm7	; xmm6=(04 05 14 15 24 25 34 35 44 45 54 55 64 65 74 75)
452
453	movdqa    xmm4,xmm1	; transpose coefficients(phase 2)
454	punpcklwd xmm1,xmm5	; xmm1=(00 01 02 03 10 11 12 13 20 21 22 23 30 31 32 33)
455	punpckhwd xmm4,xmm5	; xmm4=(40 41 42 43 50 51 52 53 60 61 62 63 70 71 72 73)
456	movdqa    xmm2,xmm6	; transpose coefficients(phase 2)
457	punpcklwd xmm6,xmm0	; xmm6=(04 05 06 07 14 15 16 17 24 25 26 27 34 35 36 37)
458	punpckhwd xmm2,xmm0	; xmm2=(44 45 46 47 54 55 56 57 64 65 66 67 74 75 76 77)
459
460	movdqa    xmm3,xmm1	; transpose coefficients(phase 3)
461	punpckldq xmm1,xmm6	; xmm1=(00 01 02 03 04 05 06 07 10 11 12 13 14 15 16 17)
462	punpckhdq xmm3,xmm6	; xmm3=(20 21 22 23 24 25 26 27 30 31 32 33 34 35 36 37)
463	movdqa    xmm7,xmm4	; transpose coefficients(phase 3)
464	punpckldq xmm4,xmm2	; xmm4=(40 41 42 43 44 45 46 47 50 51 52 53 54 55 56 57)
465	punpckhdq xmm7,xmm2	; xmm7=(60 61 62 63 64 65 66 67 70 71 72 73 74 75 76 77)
466
467	pshufd	xmm5,xmm1,0x4E	; xmm5=(10 11 12 13 14 15 16 17 00 01 02 03 04 05 06 07)
468	pshufd	xmm0,xmm3,0x4E	; xmm0=(30 31 32 33 34 35 36 37 20 21 22 23 24 25 26 27)
469	pshufd	xmm6,xmm4,0x4E	; xmm6=(50 51 52 53 54 55 56 57 40 41 42 43 44 45 46 47)
470	pshufd	xmm2,xmm7,0x4E	; xmm2=(70 71 72 73 74 75 76 77 60 61 62 63 64 65 66 67)
471
472	mov	edx, JSAMPROW [edi+0*SIZEOF_JSAMPROW]
473	mov	esi, JSAMPROW [edi+2*SIZEOF_JSAMPROW]
474	movq	XMM_MMWORD [edx+eax*SIZEOF_JSAMPLE], xmm1
475	movq	XMM_MMWORD [esi+eax*SIZEOF_JSAMPLE], xmm3
476	mov	edx, JSAMPROW [edi+4*SIZEOF_JSAMPROW]
477	mov	esi, JSAMPROW [edi+6*SIZEOF_JSAMPROW]
478	movq	XMM_MMWORD [edx+eax*SIZEOF_JSAMPLE], xmm4
479	movq	XMM_MMWORD [esi+eax*SIZEOF_JSAMPLE], xmm7
480
481	mov	edx, JSAMPROW [edi+1*SIZEOF_JSAMPROW]
482	mov	esi, JSAMPROW [edi+3*SIZEOF_JSAMPROW]
483	movq	XMM_MMWORD [edx+eax*SIZEOF_JSAMPLE], xmm5
484	movq	XMM_MMWORD [esi+eax*SIZEOF_JSAMPLE], xmm0
485	mov	edx, JSAMPROW [edi+5*SIZEOF_JSAMPROW]
486	mov	esi, JSAMPROW [edi+7*SIZEOF_JSAMPROW]
487	movq	XMM_MMWORD [edx+eax*SIZEOF_JSAMPLE], xmm6
488	movq	XMM_MMWORD [esi+eax*SIZEOF_JSAMPLE], xmm2
489
490	pop	edi
491	pop	esi
492;	pop	edx		; need not be preserved
493;	pop	ecx		; unused
494	poppic	ebx
495	mov	esp,ebp		; esp <- aligned ebp
496	pop	esp		; esp <- original ebp
497	pop	ebp
498	ret
499
500; For some reason, the OS X linker does not honor the request to align the
501; segment unless we do this.
502	align	16
503